Hi Salva and everyone,

It is ready. Cobble Flink 0.3.0-1 is now released, with several performance
improvements. Please feel free to try it out if you are interested.

I also compared the performance of Cobble and RocksDB [1]. Cobble performs
significantly better in many scenarios and is comparable in others.
Performance, however, is not Cobble’s primary differentiator. Future
development will focus on providing a unified, open, and user-friendly
storage layer for Flink.

Also cc'ing @Levani who might be interested in this.


Best,
Zakelly

[1]
https://cobble-project.github.io/cobble-flink/latest/state-backend/benchmark.html

On Tue, Jul 28, 2026 at 11:40 PM Salva Alcántara <[email protected]>
wrote:

> Looks great! Thanks a lot for sharing, Zakelly. Is it ready to go? I'd
> like to try it out / experiment with it a bit...
>
> Regards,
>
> Salva
>
> On Tue, Jul 21, 2026 at 6:17 PM Zakelly Lan <[email protected]> wrote:
>
>> Hi Flink community,
>>
>> I would like to introduce the cobble-flink project, an open-source
>> integration between Apache Flink and the Cobble storage engine[3].
>>
>> The cobble-flink aims to provide a unified storage for stream processing,
>> covering managed state, sources, and sinks. In particular, it makes Flink
>> state easier to observe, understand, and consume. More details are
>> available in the project repository [1] and documentation [2].
>>
>> By building on Cobble's remote storage and distributed snapshot
>> capabilities, cobble-flink also enables disaggregated storage, with fast
>> state rescaling when Flink job parallelism changes.
>>
>> Its capabilities are designed to work together around the same persisted
>> data:
>>
>> * Flink jobs write data through the keyed state backend or SQL sink
>> * Other Flink jobs consume it through scans or exact-key lookups
>> * Users/AI agents inspect checkpoints, savepoints, state, timers, and
>> sink snapshots through the web monitor or Java SDK
>>
>> This connects state storage, table storage, downstream consumption, and
>> inspection in one workflow. Persisted state can be displayed as named,
>> typed fields, consumed through Flink SQL, and reused in lookup joins or new
>> pipelines.
>>
>> The cobble-flink supports Flink 1.17 and later. It is licensed under
>> Apache License 2.0 and is an independent project.
>>
>> The project is currently under rapid development and is not yet
>> recommended as a mature production solution. Ideas, feedback, and
>> contributions are very welcome.
>>
>>
>> Best regards,
>> Zakelly
>>
>>
>> [1] https://github.com/cobble-project/cobble-flink
>> [2] https://cobble-project.github.io/cobble-flink/latest
>> [3] https://github.com/cobble-project/cobble
>>
>

Reply via email to